Develop Python GPU kernels with explicit shared memory and warp-level control.
npx skills add https://github.com/hung-phan/ml-skills --skill tilelang-hung-phan
Please help me install this Agent Skill. Skill: tilelang Source: https://github.com/hung-phan/ml-skills/tree/main/skills/ml-review/references/gpu-lang/tilelang Command: npx skills add https://github.com/hung-phan/ml-skills --skill tilelang-hung-phan
TileLang addresses the gap in GPU kernel development tooling by providing explicit control over shared memory, warp-level operations, and pipeline staging, compiling to CUDA/HIP via TVM/TIR.
Install TileLang and write your first kernel to perform a GEMM operation.
High-intent search queries and answers about installing and using this skill.
Write high-performance GPU kernels in Python using a DSL that provides explicit control over shared memory, warp-level operations, and pipeline staging. Code is lowered to CUDA or HIP via TVM Tensor IR for optimized hardware execution.
Yes, TVM TIR supports compiling Python GPU kernels to CUDA, HIP, WebGPU, and CPU. This lowering process translates high-level Pythonic DSL code into optimized hardware-specific instructions.
Implement multi-stage software pipelining for GPU computations by explicitly defining pipeline depth in a Pythonic DSL. This approach hides memory latency and optimizes complex tasks like FlashAttention and convolutions.
Yes, you can optimize FlashAttention and MLA computations using a Pythonic DSL. It allows fine-grained management of shared memory and warp-level primitives to achieve high performance on complex GPU operations.
Yes, TVM is required as a dependency. The framework lowers Python GPU kernels through TVM's Tensor IR to generate executable code for CUDA, HIP, WebGPU, and CPU backends.
Use a Python DSL for GPU programming to gain fine-grained control over shared memory and pipeline staging while avoiding complex low-level CUDA syntax. It lowers through TVM TIR to achieve comparable hardware performance.
hairyfCompile, migrate, and test smart contracts on TRON and EVM-compatible chains.
tensormuxPort CUDA kernels to HIP for ROCm-enabled AMD GPUs.
ovachieverExtract nuclei and cellular features from whole-slide pathology images.
hung-phanCreate high-performance GPU kernels for deep learning using Python.
